nx_nofloat_incontext_gate.nx

nx_nofloat_incontext_gate.nx -- IN-CONTEXT COPYING (CAP-NF-INCONTEXT): the basis of in-context learning. Task = ECHO: at each position predict the PREVIOUS token (tgt[i]=S[i-1]). EVERY training sequence is FRESH RANDOM (LCG), so the model CANNOT memorize sequences -- it must learn the copy OPERATION (a "previous-token attention head": attend to relative position -1 and copy its token). Then it echoes NEVER-SEEN random sequences -> generalization to unseen DATA, the foundation of in-context learning. T1 held-out echo accuracy >> chance (1/V) on fresh random sequences = the model copies arbitrary unseen content. T2 same on a DIFFERENT test seed = the generalization is robust (not seed-luck). Pure integer Q16, attention-only block (2 clm_fwd call sites). HONEST: copy-back-1 (a previous-token head), the simplest in-context copy -- not full content-based induction. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0